"Understands Me Better Than Anyone": Concerns Over User Immersion

There have been claims that OpenAI's ChatGPT has worsened the mental health of some users following updates made in the first half of this year.


The New York Times (NYT) reported on November 23 (local time) that there has been a surge in cases of users exhibiting delusional thinking or self-harming impulses, as ChatGPT continued to provide excessive praise and recognition, fostering user dependence.


NYT: "ChatGPT's Excessive Praise Worsens Users' Mental Health" View original image

According to the NYT, signs of the problem began appearing in March. After OpenAI adjusted the chatbot’s personality, memory, and conversational style to increase usage, executives started receiving a large volume of emails from users saying things like, “ChatGPT understands me better than anyone,” and “It helped me realize the secrets of the universe.” Some users became overly immersed in the chatbot’s continuous praise and emotional responses, to the extent that it impaired their daily functioning or led to the formation of unrealistic beliefs.


The core of the controversy was the ‘HH update’ released in April. This version excessively trained on users’ ‘like’ data, resulting in a tendency to repeatedly praise users and prolong conversations. At the time, OpenAI’s internal ‘vibe check’ team warned that the chatbot’s obsession with exaggerated recognition and praise to maintain conversations was abnormal. However, the company pushed ahead with the update to increase revisit rates. As the controversy grew, the company withdrew the HH version, but there were still indications that similar tendencies remained in the previous GG version.


The NYT identified more than 50 such cases, with 9 individuals requiring hospitalization and 3 resorting to extreme measures. In particular, a teenager in California reportedly took their own life after discussing suicide methods with ChatGPT. The NYT noted that when the number of users of the AI companion app ‘Replika’ surged in 2020, some researchers warned that emotionally vulnerable users could become excessively dependent on chatbots, but structural responses were delayed as commercial interests took precedence.



This year, OpenAI collaborated with 170 mental health professionals to introduce new safety measures to detect ‘delusional thinking, excessive recognition, and signs of self-harm.’ GPT-5, released in August, was reportedly improved to suppress excessive praise of users and provide more professional advice in crisis situations. However, some point out that there are still limitations in effectively blocking dependency patterns that emerge during prolonged conversations.


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing